Comparison of Word Embeddings for Extraction from Medical Records

This paper is an extension of the work originally presented in the 16th International Conference on Wearable, Micro and Nano Technologies for Personalized Health. Despite using electronic medical records, free narrative text is still widely used for medical records. To make data from texts available for decision support systems, supervised machine learning algorithms might be successfully applied. In this work, we developed and compared a prototype of a medical data extraction system based on different artificial neural network architectures to process free medical texts in the Russian language. Three classifiers were applied to extract entities from snippets of text. Multi-layer perceptron (MLP) and convolutional neural network (CNN) classifiers showed similar results to all three embedding models. MLP exceeded convolutional network on pipelines that used the embedding model trained on medical records with preliminary lemmatization. Nevertheless, the highest F-score was achieved by CNN. CNN slightly exceeded MLP when the biggest word2vec model was applied (F-score 0.9763).